DFTGL: Domain Filtered and Target Guided Learning for Few-Shot Anomaly Detection

Abstract This paper addresses cross-domain challenges in few-shot anomaly detection, where utilizing various source domains leads to diminished representations and compromised detection in the target domain. To tackle this, we propose Domain Filtering and Target-Guided Learning (DFTGL). Initially, we measure domain gaps and retain source domains with smaller disparities. We introduce a limited number of target domain samples to create an intermediate domain for better feature transfer during training. Additionally, we employ category-prior-based augmentation to refine feature distribution estimation while ensuring image registration. Experimental results demonstrate significant improvements in image-level AUROC compared to the baseline: 5.1%, 5.9%, and 4.1% (2-shot, 4-shot, and 8-shot settings) on MVTec and 6.9%, 2.1%, and 2.5% on ViSA datasets. This pioneering research effectively narrows domain gaps, enabling proficient feature transfer, and holds promise for early anomaly detection in industries like product inspection and medical diagnostics.